Human-Centric Active Perception for Autonomous Observation

David Kent, Sonia Chernova

TL;DR

This paper explores how autonomous robots (like NASA’s Astrobee) can best film humans in complex environments like the International Space Station. By comparing two mathematical frameworks—Multi-Objective MDPs and Constrained MDPs—the researchers show that while both can successfully balance filming quality with human comfort, Constrained MDPs are much easier for developers to configure, despite being significantly slower to compute.

What Problem Does This Paper Address?

In high-stakes environments like the International Space Station (ISS) or deep-sea research stations, recording human activity is vital for documentation and later analysis. Traditionally, this requires a human operator or static cameras. As robots like NASA’s Astrobee (a free-flying cube) become more autonomous, they are being tapped to act as "autonomous cameramen."

However, filming a moving human in a 3D environment is hard. The robot must maximize "observation reward" (getting a clear, high-resolution shot of the task) while balancing three conflicting factors:

  1. Safety: Avoiding collisions with the human or their workspace.
  2. Intrusion: Not annoying the human with noise (fans) or by hovering too close to their face.
  3. Efficiency: Managing limited battery life and robot movement costs.

The authors seek to find the best mathematical way to model these trade-offs so a robot can plan its movements effectively.

Methods

The authors formalize the problem as a Semi-Markov Decision Process (SMDP). Unlike a standard MDP, an SMDP allows for actions that take variable amounts of time (like moving between different modules of the ISS).

They compare two primary optimization strategies:

  • MOMDP with Scalarization: This method takes all rewards and costs and mashes them into a single number using a weighted sum. The robot then tries to maximize this one number. It is solved using backwards induction.
  • Constrained MDP (CMDP): This method picks one primary goal (maximize filming quality) and treats everything else as a "budget" or hard limit (e.g., "don't exceed 20 units of intrusion"). It is solved using Linear Programming.

To handle the unpredictability of humans, the authors model human movement as a probability distribution based on previous task data rather than assuming the human is a static point.

Datasets

The study was conducted using NASA’s Gazebo simulation of the International Space Station. While they didn't use a massive external dataset, they created three distinct "human task" scenarios for evaluation:

  1. Experiment: The human moves between three fixed stations and stays for a while.
  2. Inspection: The human moves slowly and continuously over a large surface area.
  3. Equipment Transfer: The human moves back and forth repeatedly over long distances.

For each task, they generated five different evaluation trajectories to ensure the robot's planning wasn't just "memorizing" one specific path.

Key Insights and Ideas

  • Satisficing vs. Optimizing: The biggest "aha moment" is the human-centric benefit of CMDPs. In MOMDPs, developers have to tweak abstract weights (e.g., "is a collision worth 0.4 or 0.5?"), which is unintuitive. CMDPs allow for satisficing—setting a clear budget (e.g., "use no more than 40% battery"). This is much easier for a human designer to reason about.
  • The Perching Advantage: The Astrobee can "perch" on handrails. The model accounts for this: perching saves power and allows the robot to turn off its noisy fans, reducing "intrusion" while still acting as a pan-tilt camera.
  • The Computational Wall: While CMDPs are easier to tune, they are computationally expensive. The MOMDP solved the 180-second task in 1 second, while the CMDP took a median of 237 seconds.

Key Results

  • Performance Parity: Both methods successfully produced optimal policies that approached the "perfect knowledge" baseline. In all scenarios, the robots effectively tracked the human while respecting safety and power constraints.
  • Transparency: CMDPs proved more robust when tasks changed. A weight set tuned for "Inspection" might fail on "Equipment Transfer" in an MOMDP, but a "budget" in a CMDP remains a budget regardless of the task.
  • Real-time Viability: Because the CMDP takes nearly 4 minutes to solve, it currently cannot handle "reactive" re-planning if a human does something totally unexpected. The MOMDP, however, is fast enough for real-time updates.

Limitations and Future Work

The primary limitation is the solve time for CMDPs. In a real-world scenario, if a human trips or changes their plan, the robot needs to re-calculate its path in milliseconds, not minutes.

The authors suggest future work should look into approximate CMDP solvers, such as Constrained Monte Carlo Tree Search, to get the "ease of tuning" of a CMDP with the "speed" of an MOMDP. They also aim to move from fully observable environments to those with more uncertainty (POMDPs).

Summary updated

Abstract

As robot autonomy improves, robots are increasingly being considered in the role of autonomous observation systems - free-flying cameras capable of actively tracking human activity within some predefined area of interest. In this work, we formulate the autonomous observation problem through multi-objective optimization, presenting a novel Semi-MDP formulation of the autonomous human observation problem that maximizes observation rewards while accounting for both human- and robot-centric costs. We demonstrate that the problem can be solved with both scalarization-based Multi-Objective MDP methods and Constrained MDP methods, and discuss the relative benefits of each approach. We validate our work on activity tracking using a NASA Astrobee robot operating within a simulated International Space Station environment.
BibTeX
@inproceedings{kent2020human,
  author = {David Kent and Sonia Chernova},
  title = {{Human-Centric Active Perception for Autonomous Observation}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {2020},
  url = {https://openreview.net/forum?id=QYv_huzzwFV},
}

Loading related content...

Figure